With the fast-growing research activities, more biomedical papers are being published in thousands of journals worldwide. For example, PubMed Central (PMC) has 5.2 million papers and 7409 journals covering biomedical and life sciences [1]. Although these publications play a major role in disseminating research outcome, the growth of journal publications imposes a challenge for selections of appropriate publication venues. It is vital that authors submit to the right journal that meets the journal scope and provides sound reviews. It is equally important that they reach their intended audience and obtain a large number of citations [2]. However, researchers are unfamiliar with all the journals related to their work for choosing the most suitable one for submitting a paper. Moreover, different publication scopes of journals and research interests of reviewers and editors may affect the decision of a submitted manuscript. If the submitted paper cannot meet the interests of a publication venue and its editors and reviewers, it may lead to rejection, delay, or less readership. An appropriate recommender system can help solve this problem.

Recommender systems have been proven to serve as an effective method for decision making in many areas such as music, movies, and information media choices [3-6]. The well-known techniques of recommender systems are content-based recommendation [7,8], collaborative filtering recommendation [4,9], and hybrid recommendation [6,10]. Content-based recommender systems recommend an item to a user based on a description of the item. Collaborative filtering methods and hybrid methods may outperform the content-based recommendations by applying user data, if available. However, after the user privacy issue of Facebook in 2018 and the introduction of European Union General Data Protection Regulation, user data are no longer easy to obtain. Moreover, in many domains, especially in material recommendation, there are no user data available for collaborative filtering methods at the beginning [11], which is regarded as a cold-start problem. Content-based recommendations do not need any user information and are more suitable for solving these problems [12].

Based on the content-based recommendation strategy, several attempts have been made to create recommender systems for medical applications and scientific literature. Using geotagged mobile search logs, Agarwal et al [13] adopted a Random Forest model to predict medical visits. Using topic, writing style, author information, citation information, abstract, and title as information items, latent Dirichlet allocation [14] and k-nearest neighbor [15] were used to classify the scientific literature for recommendation [2,12,16,17]. Luong et al [18] used the coauthors’ network as advanced information to recommend a publication venue. Beel et al [19] conducted a literature survey on recommender systems, exploring their methods, evaluation measurements, and datasets. For most of these recommender systems, the high-dimensional and sparse matrix computation is a critical problem [20].

Because of the mismatches caused by ambiguity in text comparisons, the content-based recommendation approach may cause a high error rate [21]. Recently, due to the ability of discovering intricate structures and deep semantics in high-dimensional data, deep learning methods have succeeded in many areas and recently been proposed to build recommender systems for both collaborative filtering and content-based approaches. Hinton et al [22] proposed restricted Boltzmann machines for modeling tabular or count data as a collaborative filtering model on the Netflix data set. McAuley et al [23] proposed an image-based recommendation, which adopted a deep learning model to extract image features. Van den Oord et al [24] applied a deep convolutional neural network (CNN) to predict latent factors from music audios for music recommendation. Wang et al [11] proposed a collaborative deep learning model to jointly perform deep representation learning for the content information and collaborative filtering of a rating matrix. However, to the best of our knowledge, these deep learning techniques have not been used in any biomedical literature recommender system.

Most current venue recommendation studies focus on computer science and technology, but not on the biomedical field. Biomedical sciences are highly interdisciplinary and often link to engineering, medicine, biology, physics, psychology, etc, thereby serving more journals and more diverse topics than any other field. Hence, the development of a recommender system is more essential and challenging for the biomedical sciences than any other discipline. Furthermore, previous recommender systems were based on shallow machine learning methods and social networks. They were generally keyword-based methods and did not take semantics into account. In addition, the few existing systems only focus on journals under a certain organization, such as Elsevier, IEEE, and Springer, instead of PubMed.

In contrast to our previous study on computer science publication recommendations using conventional machine learning approaches [12], we proposed a deep learning–based recommender system for biomedical publication venues, named Pubmender. Due to the copious vocabulary of biomedical literature, the traditional vector space model can lead to high-dimensional and sparse problems. To address this issue, dimensionality reduction methods are needed before learning the pattern. Moreover, initializing text matrix by pretrained word embedding is more beneficial for training neural networks than random initialized embedding [25]. Accordingly, we applied a word2vec model for our study instead of using the conventional vector space model employed in our previous publication venue recommender system. In addition, deep learning models are able to learn multiple-level abstract representations of data with syntactic and semantic information, since more abstract concepts can be constructed with multiple processing layers [26]. We applied the deep learning approach to provide recommendations of journals for biomedical researchers. Unlike shallow learning, the state-of-the-art embedding method and deep CNN in Pubmender were trained from 837,882 papers in 1130 biomedical journals. This method can help researchers find a variety of choices, without being limited to their own knowledge of journals.

Figure 1 shows the architecture and workflow of our Pubmender system. It consists of user interface, data preprocessing, abstract representation, classification, and ranking phase.

The user interface obtains the input data (an abstract submitted by a user) and presents the recommendation results to the user. The data acquirement is followed by data preprocessing and information extraction. At the start of our deep learning model, the abstract representation phase converts an abstract to a vector. The original abstract vector is a concatenation of pretrained word vectors. Subsequently, deep CNN is applied to train the model to achieve high-level abstract representation. A three-layer fully connected network with a softmax operation is applied to classify papers based on the obtained abstract vectors. The recommendation list of the top N journals obtained from the ranking phase is presented to the user.

The data were downloaded from the File Transfer Protocol service of PubMed Central (PMC) [27], containing 1,534,649 papers. Based on the journal list of PMC, we selected normal journals deposited under full participation or the US National Institutes of Health portfolio mode, excluding records labeled “Predecessor,” “No New Content,” and “Now Select.” Papers from Jan 2007 to Apr 2017 were selected. Papers with no abstracts or with fewer than 200 characters in abstracts were deleted. Journals containing fewer than 100 papers were also removed. Finally, 880,165 papers in the XML format from 1130 journals were used in our study.

Each PMC file is a semistructured XML document and contains various tags, such as <title>, <abstract>, and <issn>. We extracted the content in <abstract>, <ISSN>, and <pub-date> fields from the raw XML files. Then, pissn and eissn in the ISSN field were replaced by “LocatorPlus ID,” which is the unique identification for a journal in the US National Library of Medicine catalog. After extraction, each abstract was stored in a corresponding file. Natural Language Toolkit was adopted to operate word segmentation [28].

In Pubmender, the recommendation task is formulated into a multilabel classification problem, where the text representation and classification methods are critical. For abstracts, we originally embedded abstracts with pretrained word vectors. Thereafter, the original embeddings were fed into CNN to achieve more abstract representation as explained below.

Let v_i ∈R^k be the k-dimensional word vector corresponding to the i-th word in the abstract A. An original representation of A is represented as a matrix V={v₁,…,v_m}^T, which is the concatenation of the words’ vectors. Due to the different sizes of abstracts, we set m as the maximum count of words in an abstract. A padding operation with zeros was adopted for input with fewer than m words in an abstract and a tail truncation operation for more than m words. The vectors of words adopt pretrained vectors using word embedding and are induced from the PubMed abstracts and PubMed Central full text. The word2vec tool [29] was adopted for word embedding using the skip-gram model with a window size of h, hierarchical softmax training, and a frequent word subsampling threshold to create k-dimensional vectors. Word vectors are initialized by zeros if they are not in the pretrained vocabulary. Finally, the representation of an abstract is matrix V with a dimensionality of m*k. It was used as the input to feed to the next step. To achieve more abstract and semantic features, we adopted CNN to extract semantic information.

Figure 2 shows the structure of our deep CNN model. There are three convolutional and max-pooling layers in CNN, one fully connected layer, one hidden layer, and one softmax layer for classification. For an abstract, A(w₁,w₂,...,w_n) with w_i represents the i-th word and v_i∈R^k is the k-dimensional word vector corresponding to word w_i. The abstract is represented as v_1:m=v₁⨁v₂⨁...⨁v_m (1), where ⨁ is the concatenation operator, m is the maximum length of abstracts (a scalar), and v_i:i+j refers to the vector of concatenation of the words w_i,w_i+1,…,w_i+j. The first convolutional layer performs as a one-dimensional convolution operation on sliding windows of h₁ words to produce a phrase feature. For example, a feature c_ji is generated from a window of words v_i:i+h1–1 by c_ji=g(f_j▪v_i:i+h1–1+b₁) (2). Here, b₁∈R is a bias term and g is a nonlinear function such as rectified linear unit (ReLu). f_j∈R^{k × h₁} is the j-th convolutional kernel, whose shape is k × h₁, where k is the dimension of word vectors and h₁ is the window size. This kernel is applied to each possible window of words in the abstract {v_1:h1,v_2:h1+1,...,v_m–h1+1:m} to produce a feature map C_j=[c_j1, c_j2...,c_j,m–h1+1] (3) with C_j∈R^m–h₁+1.

If there are r₁ convolutional kernels, then is the result of the first convolution operation on V. The pooling operation is then carried out on C⁽¹⁾. Its function is to progressively reduce the spatial size of the representation to extract the key features and reduce the number of dimensions in the network. The pooling layer operates independently on every depth slice of the input and resizes it spatially, using the max-pooling operation [30] in every two-unit window for each C_j⁽¹⁾. P⁽¹⁾, described below, is the result of the max-pooling operation: (4), where j is the j-th filter of the convolutional operation (5).

The second and third convolutional and pooling layers work the same as Equations (2) and (5). Following the three convolutional and pooling operations is the fully connected layer. Here, the input is represented with a more abstract feature , where r₃ is the number of third-layer convolutional filters. The three convolutional and pooling operations indicate a phrase-level feature, a sentence-level feature, and an abstract-level feature. The algorithm of abstract embedding is listed in Textbox 1.

A fully connected softmax layer is the last layer of Pubmender. Given the training sample, A, where T is the number of possible labels, z is the class score for the sample, and the estimated probabilities S_j∈[0,1) for each label j∈{1,2…T} the softmax formula is:

We trained the entire model by minimizing the cross-entropy error defined as (7), where Y is the true classification output. This is a one-hot encoding of size T, where all elements except one are 0, and one element is 1. This element marks the correct class for the data classified. We employed the optimizer Adam to learn the model parameters, which is a variant of stochastic gradient descent [31].

After data preprocessing, there were 880,165 preprocessed papers from PMC in 1130 open-access journals from Jan 2007 to Apr 2017. The “LocatorPlus ID” assigned to each journal by PMC is regarded as the classification label of a paper. We generated four data sets based on these papers. The first data set included all papers from 2007 to 2016, which was used to choose the feature representation method and train the prediction models. Papers in 2017 formed the second data set, which was used as the test set to verify Pubmender’s performance. The last two datasets chose papers from publications in Elsevier and Springer from 2017, which were used to compare our Pubmender with Journal Finder and Suggester. The statistics of the first dataset are described in Table 1.

One of the biggest challenges of these datasets is that the data distribution is highly imbalanced. In the first dataset, 60 journals published more than 2000 papers, while 740 journals published fewer than 400 papers. The number of papers in “PLOS One” was 153,608, which is larger than the number in other journals, based on its extensive and comprehensive scope. “Scientific Reports” ranked second, with 37,864 papers published, while “Horticulture Research” only published 100 papers. The average paper count was 741, and 934 journals had fewer than that number.

For CNN, three convolutional and three pooling operations were adopted. The pretrained word vectors generated by word2vec, available from Evex [29], were used. The window size h was 5 and threshold was 0.001. The dimension of a pretrained vector was 200, that is, k=200. The length of abstract had a fixed size m, which is the maximum number of the words that most abstracts contain. Table 2 shows the word statistic details of the papers. With the statistics, only 43,328 of 837,882 papers (5%) contained abstracts with more than 350 words. Therefore, we chose m=350. A zero-padding operation was applied for abstracts with fewer than 350 words, together with a tail-truncation operation for abstracts containing more than 350 words.

The convolutional operation parameters are listed in Table 3. The activation function adopted the ReLU function. In the pooling layer, the size of max-pooling filters was two, applied with a stride of two down samples. The parameters of the following layers (pooling layers) had the same parameter settings. Normalization and dropout strategies were used in the fully connected layer. Rate of dropout was 0.2 and L₂ normalization was adopted.

Journal Finder is provided by Elsevier for recommending Elsevier journals [35]. There are 45 Elsevier journals in our dataset. Five of them with more papers were selected. The paper counts were 582 for Medicine, 193 for Data in Brief, 124 for NeuroImage: Clinical, 117 for Redox Biology, and 87 for Preventive Medicine Reports.

Elsevier’s Journal Finder requires input of the title and abstract of submitted paper, and fields of research. The titles and abstracts were extracted from XML files and then fed into Journal Finder. We chose fields of research in “Engineering,” “GeoSciences,” “Life and Health Science,” and “Chemistry.” The results are listed in Table 8 and show that our system is much better than Journal Finder.

Journal Suggester [36], recommends journals published by Springer. Journal Suggester also requires input of the title and abstract, and field of research. We chose “Biomedicine” as the field of research. There are 14 journals from Springer in our dataset, and seven of them were chosen for comparison based on a significant number of papers in each journal — Cell Death & Disease, Malaria Journal, Nanoscale Research Letters, Nature Communications, Parasites & Vectors, Scientific Reports, and Trials. Each journal chose the top 100 papers according to the size of the XML files. The results are listed in Table 9. Again, our system was much better than Journal Suggester.

For biomedical publications, our Pubmender system is the first recommender system with word embedding and deep learning models. It achieves 87.0%, 22.9%, and 196.0% higher accuracy than recommender systems on MAS, ACM, and CiteSeer, respectively. In addition, the experiment results also revealed that the accuracy of our system was superior to that of Journal Finder and Journal Suggester. Our web service is freely available online [37].

Because no paper has been published about biomedical venue recommendations, we cannot perform any exact comparison with previous work. However, some publishers provide tools to help authors choose suitable journals. We chose two tools provided by Elsevier and Springer for comparison. The first one is Journal Finder provided by Elsevier for recommending journals of Elsevier. Pubmender achieved a much higher accuracy than Journal Finder on four metrics. For example, on acc@1, the accuracy of our system reached 0.62 and Journal Finder was given an accuracy rating of 0.05, with 1140% improvement; on acc@10, the accuracy of our system reached 0.84, which is 546% higher than that of Journal Finder. Pubmender also significantly outperformed another tool, Journal Suggester.

In this study, we proposed a biomedical publishing venue recommender system—Pubmender. In this system, an abstract is first represented by a vector using the composition of pretrained word vectors. Subsequently, a deep CNN architecture is designed to represent and classify the submitted abstract. The original vectors are converted into more abstract feature vectors containing semantic information using deep CNNs, which overcome the sparse high-dimensional problem.

The experimental results showed that our proposed system achieves more successful performance than that of MAS, ACM, CiteSeer, Journal Finder, and Journal Suggester. Even for journals containing a small number of abstracts, the performance of Pubmender was satisfactory, because Pubmender’s high-level representation method catches more semantic and structural information from the abstract.